What this category of data looks like
The data for this category comes from the centralized treasury payment system, unified budget management system, electronic invoice platform, and contract management system. The update frequency is daily. Daily documents include payment applications, fund use plans, invoices, contracts, acceptance documents, reimbursement approval forms, and payment vouchers. Fields include budget indicators, economic classifications, payment amounts, approval nodes, contract numbers, and more. The data has both structured fields and unstructured text, with some documents containing embedded images. Cross-system data requires association and matching. There is no fixed unified format. Field consistency across data from different systems must be ensured.
What constraints do these characteristics impose on knowledge base retrieval and recall
Daily updated data requires the knowledge base to match the daily update rhythm. This avoids using expired data that affects audit accuracy. The multi-source and heterogeneous document structure requires retrieval to support both structured field matching and unstructured text recall. This ensures coverage of core fields such as budget indicators and economic classifications. The high-risk nature of single payment audits requires retrieval results to accurately match core fields. This prevents issues such as payment returns or incorrect fund disbursement caused by mismatches. Cross-system data association requires retrieval to link associated information of contracts, invoices, and payment vouchers. This ensures a complete audit chain.
How to set up the configuration
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matching core fields such as budget indicators and economic classifications requires a high similarity threshold. This ensures recall results meet the accuracy requirements of fiscal audits |
limit | Top 10–15 entries | Single payment audits need to cover multiple types of documents including contracts, invoices, and acceptance documents. 10–15 entries balances retrieving sufficient associated information and avoids missing key audit basis |
searchMode | Full-text retrieval + structured field retrieval | Documents in this category have both structured fields and unstructured text. The mixed retrieval mode can match fields such as budget indicators and the text content of invoices and contracts at the same time |
usingReRank | Enabled | Multi-source recall results need to be reordered by relevance. This ensures the most relevant audit basis is displayed first, improving audit efficiency |
datasetSearchUsingExtensionQuery | Enabled | For documents such as payment applications and reimbursement approval forms, expand query terms. This supplements retrieval coverage of associated fields such as budget indicators and economic classifications |
collectionIds | Bind the dedicated knowledge base collection for this category | Limit the retrieval scope to the knowledge base of the centralized treasury payment and unified budget management systems. This avoids interference from irrelevant data on audits |
Three common mistakes
- Phenomenon: Retrieval results fail to match documents related to the corresponding budget indicators. Cause: Structured field retrieval mode is not enabled. Relying only on full-text retrieval cannot accurately locate core fields such as budget indicators and economic classifications.
- Phenomenon: Embedded images in PDF documents uploaded to the knowledge base cannot be displayed properly. Cause: OCR indexing configuration for PDF embedded images is not configured. The text in images is not included in the retrieval knowledge base.
- Phenomenon: The number of retrieval results returned in a single query exceeds the set range. Cause: The
limitparameter is not configured correctly, or too many irrelevant knowledge base collections are bound tocollectionIds. This results in recall of documents not belonging to this category.
How to confirm the configuration is correct
- Import a single simulated payment application data. Verify whether the retrieval results include documents matching the corresponding budget indicators and economic classifications. Check whether the
similaritythreshold is within the set range. - Upload a PDF contract document containing embedded images. Check whether the images can be displayed normally in the knowledge base. Confirm that the OCR indexing configuration has taken effect.
- View the number of retrieval results returned. Confirm that the number is within the interval set by
limit. - Check the logs of the RAG retrieval link in the workflow. Confirm that
usingReRankis enabled, and the retrieval results are sorted by relevance.
The values given here are common starting points and should be measured against the reader's own samples.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-28.